JP2010271712A - Sound data processing device and sound data processing method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To process sound data including a signal sound so as to reduce noise to be caused by a mechanical device. <P>SOLUTION: A sound data processing device includes: an operating state acquisition part 101 acquiring an operating state of the mechanical device; a sound data acquisition part 103 acquiring sound data corresponding to the acquired operating state; and a database 105 storing various kinds of operating states of the mechanical device in unit time and associated sound data as a template. The sound data processing device also includes: a database retrieval part 107 retrieving a template of the operating state nearest to the acquired operating state from the database 105, and a template subtraction part 109 subtracting sound data of the template of an operating state nearest to the acquired operating state from the acquired sound data to reduce the noise to be caused by the mechanical device. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a sound data processing device and a sound data processing method for processing sound data including a signal sound so as to reduce noise generated by a mechanical device such as a robot.

  For example, a robot that performs automatic speech recognition while operating needs a function of suppressing noise caused by the operation of the robot itself, that is, self-noise (Patent Document 1, etc.).

  In general, sound source identification and sound source separation for noise reduction have been studied, but they are not directed to automatic speech recognition under self-noise (Non-Patent Document 1, etc.). Also, conventional noise reduction methods such as spectral subtraction do not actually work well in many cases (Non-Patent Document 2, etc.). A method of performing spectral subtraction for each operation is also known, but it is practically impossible to deal with many types of operations. Furthermore, since the noise source of the robot's self-noise exists in the near field, the performance of the conventional far-field noise reduction method is significantly degraded.

  Thus, a sound data processing apparatus and method for processing sound data including signal sound so as to efficiently reduce noise generated by a mechanical device such as robot self-noise has not been developed.

JP 2008-122927 A

I. Hara, F. Asano, H. Asoh, J. Ogata, N. Ichimura, Y. Kawai, F. Kanehiro, H. Hirukawa, and K. Yamamoto, “Robust speech interface based on audio and video information fusion for humanoid HRP-2 ”in Proc. IEEE / RSJ International Joint Conference on Robots and Intelligent Systems (IROS), pp.2404-2410, (2004) S. Boll, “Suppression of Acoustic Noise in Speech Using Spectral Subtraction”, in IEEE Transactions on Acoustics, Speech, and Signal Processing, vol.ASSP-27, No.2, (1979)

  Therefore, there is a need for a sound data processing device and a sound data processing method for processing sound data including signal sound so as to efficiently reduce noise generated by a mechanical device such as robot self-noise.

  A sound data processing apparatus according to one aspect of the present invention is a sound data processing apparatus that processes sound data including a signal sound so as to reduce noise generated by a mechanical device. The sound data processing device according to this aspect includes an operation state acquisition unit that acquires an operation state of the machine device, a sound data acquisition unit that acquires sound data corresponding to the acquired operation state, and the machine device in unit time And a database for storing various operation states and corresponding sound data as templates. The sound data processing apparatus according to the present invention includes a database search unit that searches the database for an operation state closest to the acquired operation state from the database, and an operation state that is closest to the operation state acquired from the acquired sound data. A template subtracting unit that subtracts the sound data of the template to reduce noise generated by the mechanical device.

  In the sound data processing device according to the present aspect, the template of the operation state closest to the acquired operation state among the templates representing the sound data corresponding to the various operation states of the machine device per unit time stored in the database. Sound data is used as a predicted value of the sound data of the acquired operation state. Therefore, by subtracting the predicted value from the acquired sound data, it is possible to efficiently reduce noise generated by the mechanical device according to the operating state.

  A sound data processing device according to an embodiment of the present invention includes a multi-channel noise reduction unit that reduces noise based on multi-channel sound data, an output of the template subtraction unit based on the acquired operation state, and the An output selection unit that selects one of the outputs of the multi-channel noise reduction unit.

  According to the present embodiment, it is possible to select an output having a higher noise reduction effect among the outputs of the template subtraction unit and the multi-channel noise reduction unit according to the acquired operation state.

  In the sound data processing device according to one embodiment of the present invention, the mechanical device is a robot.

  According to the present embodiment, it is possible to efficiently reduce noise generated by the robot according to the operation state of the robot.

  In the sound data processing apparatus according to one embodiment of the present invention, the operation state of the robot is represented by data on the angle, angular velocity, and angular acceleration of the joint motor.

  According to the present embodiment, the operation state of the robot can be easily and reliably grasped by using the data of the angle, angular velocity, and angular acceleration of the joint motor.

  In the sound data processing device according to one embodiment of the present invention, the template of the operation state closest to the acquired operation state is determined by the distance in the three-dimensional space constituted by the motor angle, angular velocity and angular acceleration data. .

  According to the present embodiment, by using the distance in the three-dimensional space, it is possible to easily and reliably determine the template having the operation state closest to the acquired operation state.

  In the sound data processing apparatus according to one embodiment of the present invention, the sound data of the template is represented by a frequency spectrum.

  According to the present embodiment, sound data can be efficiently stored by representing it by a frequency spectrum.

  The sound data processing device according to one embodiment of the present invention further includes a database creation unit that creates the database by collecting various operation states of the mechanical device in unit time and sound data corresponding to the operation states. Yes.

  According to this embodiment, the work load for creating a database is greatly reduced.

  A sound data processing method according to one aspect of the present invention is a sound data processing method for processing sound data including a signal sound so as to reduce noise generated by a mechanical device. The sound data processing method according to this aspect includes a step of acquiring an operation state of the mechanical device and a step of acquiring sound data corresponding to the acquired operation state. In the sound data processing according to this aspect, the sound data of the template in the operation state closest to the acquired operation state is searched from the database that stores the various operation states of the mechanical device in the unit time and the corresponding sound data as templates. And subtracting the sound data of the template in the operation state closest to the acquired operation state from the acquired sound data to obtain an output with reduced noise generated by the mechanical device.

  In the sound data processing method according to this aspect, the template of the operation state closest to the acquired operation state among the templates representing the sound data corresponding to various operation states of the mechanical device in the unit time stored in the database is stored. Sound data is used as a predicted value of the sound data of the acquired operation state. Therefore, by subtracting the predicted value from the acquired sound data, it is possible to efficiently reduce noise generated by the mechanical device according to the operating state.

  A sound data processing method according to an embodiment of the present invention is configured to output either noise-reduced output by template subtraction or noise-reduced output using multi-channel sound data based on the acquired operation state. The method further includes a step of selecting.

  According to the present embodiment, the output with the higher noise reduction effect among the output with reduced noise by template subtraction and the output with reduced noise using multi-channel sound data, depending on the acquired operating state. Can be selected.

It is a figure which shows the structure of the template subtraction noise reduction part of the sound data processing apparatus by one Embodiment of this invention. It is a figure which shows the structure of the sound data processing apparatus by one Embodiment of this invention. It is a figure which shows the structure of a database. It is a flowchart which shows the procedure which produces a database. It is a flowchart which shows the procedure of the noise reduction using template subtraction. It is a figure which shows the relationship between SN ratio (signal-to-noise ratio) and word correct rate (WCR) in case there exists noise of "head movement". It is a figure which shows the relationship between SN ratio (signal-to-noise ratio) and word correct rate (WCR) in case there exists noise of "arm movement". It is a figure which shows the relationship between SN ratio (signal-to-noise ratio) and a word correct rate (WCR) in case there exists noise of "head movement" and "arm movement".

  FIG. 1 is a diagram illustrating a configuration of a template subtraction noise reduction unit 100 of a sound data processing device according to an embodiment of the present invention. The mechanical device in this embodiment is a robot. The template subtraction noise reduction unit 100 includes an operation state acquisition unit 101 that acquires an operation state of the robot, a sound data acquisition unit 103 that acquires sound data including a signal and noise, and various operation states of the mechanical device in unit time. A database 105 that stores corresponding sound data as a template and a database creation unit 111 that creates a database are provided. As described above, the template represents sound data (acoustic signal) generated in various operation states of the robot for a unit time. The template subtraction noise reduction unit 100 searches the database 105 for a template having the operation state closest to the acquired operation state, and the operation state closest to the operation state acquired from the acquired sound data. And a template subtracting unit 109 that subtracts the sound data of the template to reduce noise generated by the mechanical device.

  The operation state acquisition unit 101 is connected to the angle sensor of the joint motor of the robot and acquires angle data from the angle sensor. The operation state of the robot is represented by the angle, angular velocity, and angular acceleration of each joint motor of the robot. The motion state acquisition unit 101 performs differential processing on the acquired angle data to obtain angular velocity data and angular acceleration data.

  The sound data acquisition unit 103 is connected to the microphone 201 installed in the robot, and acquires sound data collected by the microphone 201. The sound data acquisition unit 103 is, for example, MCRA (I. Cohen, “Noise Estimation by Minima Controlled Recursive Averaging for Robust Speech Enhancement, in IEEE Transactions on Acoustics, Speech, and Signal Processing, vol. ASSSP-27, No. 2, (1979)) or the like may be provided.

  The data structure of the database 105, the function of the database creation unit 111, the function of the database search unit 107, and the function of the template subtraction unit 109 will be described in detail later.

  FIG. 2 is a diagram illustrating a configuration of the sound data processing device 150 according to an embodiment of the present invention. The sound data processing device 150 includes a template subtraction noise reduction unit 100 described with reference to FIG. 1 and a conventional geometric source separation method (Geometric Source Separation, GSS, LC Parra and CV Alvino, “Geometric Source Separation * Multi-channel noise reduction unit 121 that reduces noise using Merging Convolutive Source Separation with Geometric Beanforming ”, in IEEE Trans. Speech Audio Process., Vol. 10, No. 6, pp. 352-362, (2002)) , Sound feature extraction units 123 and 125 that extract sound features from the noise-processed sound data, an output selection unit 127 that selects one of the output of the template subtraction noise reduction unit 100 and the output of the multi-channel noise reduction unit 121 A speech recognition unit 129 that performs speech recognition using the output output.

  The multi-channel noise reduction unit 121 acquires sound data from eight microphones 201... 203 installed on the head of the robot, specifies a sound source position, and performs sound source separation using the specified sound source position. After filtering (Ephraim and D. Malah, “Speech enhancement using a minimum mean-square error short -time spectral amplitude estimator”, in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp.901- 904, (2002)). The post-filtering process reduces, for example, stationary noise, such as background noise, and non-stationary noise caused by leakage energy between the output channels of individual sound source separation stages. The multi-channel noise reduction unit 121 may be realized by any configuration other than the above-described configuration that uses multi-channels that can separate sound sources having directions.

  The output selection unit 127 receives information on the operation state of the robot from the operation state acquisition unit 101 of the template subtraction noise reduction unit 100, and based on the information, of the template subtraction noise reduction unit 100 and the multi-channel noise reduction unit 121, The output for reducing noise more efficiently in the operating state is selected, and the output is sent to the voice recognition unit 129. The relationship between the operation state of the robot and the performance of the template subtraction noise reduction unit 100 and the multi-channel noise reduction unit 121 will be described later.

  FIG. 3 is a diagram showing the structure of the database 105.

  FIG. 4 is a flowchart showing a procedure for creating the database 105.

  When creating the database 105, the robot executes a sequence of continuous movements consisting of a large number of movements while providing a pause shorter than 1 second between individual movements, and the movement state acquisition unit 101 moves to a movement state. And the sound data acquisition unit 103 acquires sound data. “Arm movement” moves the entire arm randomly in the reach space, “Leg movement” performs stepping and short-distance walking, and “Head movement” rotates the head randomly (up and down [-30 ° 30 °], bearing [-90 ° 90 °]).

  In step S1010 of FIG. 4, the operation state acquisition unit 101 acquires the operation state of the robot for a predetermined time.

及び角加速度

によって表される。ロボットの関節の数をＪとすると、動作状態を表す特徴ベクトルは以下のようになる。

ここで、ｋは、時刻を表す。角度θ、角速度

及び角加速度

の値は、５ミリ秒ごとに取得し、[-1,1]に規格化する。 As described above, the robot operation state is determined by the angle θ and the angular velocity of each joint motor of the robot.

And angular acceleration

Represented by If the number of joints of the robot is J, the feature vector representing the motion state is as follows.

Here, k represents time. Angle θ, angular velocity

And angular acceleration

The value of is obtained every 5 milliseconds and normalized to [-1,1].

  Specifically, the robot uses two motors for head movement, five motors for each leg movement, and four motors for each arm movement. Thus, since 20 motors are used in total, J is 20.

ここで、ｋは、時刻を表し、Ｆは周波数の範囲を表す。該周波数の範囲は、０ｋＨｚ−８ｋＨｚを２５６に区分したものである。音データは、１０ミリ秒ごとに取得する。上述のように、音データ取得部１０３は、たとえば、ＭＣＲＡを使用して背景ノイズを低減した後のデータを、モータ・ノイズに対応する音データとして使用してもよい。 In step S1020 of FIG. 4, the sound data acquisition unit 103 acquires sound data for a predetermined time. Specifically, sound data corresponding to the above-described robot operation state, that is, sound data corresponding to motor noise is acquired and represented by the following frequency spectrum.

Here, k represents time and F represents a frequency range. The frequency range is obtained by dividing 0 kHz-8 kHz into 256. Sound data is acquired every 10 milliseconds. As described above, the sound data acquisition unit 103 may use, for example, data after reducing background noise using MCRA as sound data corresponding to motor noise.

を受け取り、音データ取得部１０３から、動作状態に対応する音データの周波数スペクトル

を受け取る。 In step S1030 of FIG. 4, the database creation unit 111 receives a feature vector representing the operation state from the operation state acquisition unit 101.

The frequency spectrum of the sound data corresponding to the operating state from the sound data acquisition unit 103

Receive.

  A time tag is attached to the frequency vector of the feature vector representing the operating state and the sound data. Therefore, a template is generated by combining the feature vector and the frequency spectrum having the same time tag. The database 105 shown in FIG. 3 is created as a set of templates generated in this way. As described above, since the period of the feature vector representing the operation state is 5 milliseconds, the period of the template is also 5 milliseconds.

  In the present invention, a large number of templates representing sound data (acoustic signals) generated in various operation states of a robot for a unit time (5 milliseconds in the above example) are collected to construct a database. After that, during the robot operation, the sound data of the template in the operation state closest to the operation state acquired at a certain time among the templates stored in the database is predicted, and the sound data of the operation state acquired at that time is predicted. Use as a value. As described above, the present invention provides a set of a large number of templates representing sound data (acoustic signals) generated in various operation states of the robot in unit time (5 milliseconds in the above example). It is based on the idea that it can cope with the prediction of data. This idea is based on the following assumptions.

1) The motor noise at that time depends on the angle, angular velocity and angular acceleration of the specific motor.
2) At any point in time, a similar noise frequency spectrum is produced by a combination of similar motion states (joint states).
3) The superposition of a single joint motor noise at any time is equal to the total noise at that time.

  In step S1040 of FIG. 4, it is determined whether or not sufficient templates are stored in the database 105. For example, after storing templates in database 105 for a sequence of movements including a series of “arm movements”, a series of “leg movements”, and a series of “head movements”, it is determined that sufficient templates have been stored. Also good.

  According to the above procedure, a database that can be used for prediction of sound data in all operating states of the robot can be easily created.

  FIG. 5 is a flowchart showing a noise reduction procedure using template subtraction.

  In step S2010 of FIG. 5, the operation state acquisition unit 101 acquires the operation state (feature vector) of the robot.

  In step S2020 of FIG. 5, the sound data acquisition unit 103 acquires sound data.

  In step S2030 of FIG. 5, the database search unit 107 receives the acquired operation state (feature vector) from the operation state acquisition unit 101, and searches the database 105 for a template having the operation state closest to the acquired operation state. To do.

と表し、取得された動作状態の特徴ベクトルを

と表す。そうすると、取得された動作状態に最も近い動作状態のテンプレートを検索することは、３Ｊ次元のユークリッド空間の距離

が最も小さくなる特徴ベクトル

を有するテンプレートを求めることに相当する。 Here, if the number of joints of the robot is J, the feature vector of the motion state corresponds to a point in the 3J-dimensional space. A feature vector of an operation state of an arbitrary template in the database 105

And the obtained feature state vector

It expresses. Then, searching for the template with the operation state closest to the acquired operation state is the distance in the 3J-dimensional Euclidean space.

Is the smallest feature vector

Is equivalent to obtaining a template having

で表し、テンプレートの周波数スペクトルを

で表すと、有効な信号の周波数スペクトル

は、以下の式で求められる。

テンプレートの周波数スペクトル

は、モータ・ノイズの予測値であるので、有効な信号の周波数スペクトル

は、残余のモータ・ノイズを含む。そこで、テンプレート減算を以下の式によって行なってもよい。

In step S2040 of FIG. 5, the template subtraction unit 109 receives the acquired sound data from the sound data acquisition unit 103, and receives the frequency spectrum of the searched template from the database search unit 107. Next, the template subtraction unit 109 subtracts the frequency spectrum of the template, which is a predicted value of motor noise, from the frequency spectrum of the acquired sound data. The frequency spectrum of the acquired sound data

And the frequency spectrum of the template

Is the frequency spectrum of the valid signal.

Is obtained by the following equation.

Template frequency spectrum

Is the predicted motor noise, so the frequency spectrum of the valid signal

Includes residual motor noise. Therefore, template subtraction may be performed by the following equation.

とする。 Here, α is an overestimation coefficient and makes it possible to strike a balance between perceptual signal distortion and noise reduction level. Β is the lower limit of the spectrum and reduces the influence of sharp peaks and valleys in the frequency spectrum. As an example

And

として、

が成立する場合に、テンプレート減算ノイズ低減部１００の出力を選択し、それ以外の場合には、多チャンネルノイズ低減部１２１の出力を選択してもよい。頭部の運動が存在する場合に、テンプレート減算ノイズ低減部１００の出力を選択する理由は後で説明する。関係式（５）において、０の代わりに正の定数εを使用する理由は、頭の運動は停止したが、関節モータの角度センサが、非常に小さな位置の差の信号を送る状況に対応するためである。出力を選択した後、出力選択部１２７は、選択した出力を音声認識部１２９へ送り、音声認識部１２９が音声認識を行なう。 In step S2050 of FIG. 5, the output selection unit 127 receives information on the robot operation state from the operation state acquisition unit 101 of the template subtraction noise reduction unit 100, and based on the information, the template subtraction noise reduction unit 100 and the multi-channel Any output of the noise reduction unit 121 is selected. The speed of the movement to shake the head up and down, left and right or tilt the head

As

Is established, the output of the template subtraction noise reduction unit 100 may be selected. In other cases, the output of the multi-channel noise reduction unit 121 may be selected. The reason for selecting the output of the template subtraction noise reduction unit 100 when there is head movement will be described later. The reason for using a positive constant ε instead of 0 in relation (5) corresponds to the situation where the head motion has stopped but the angle sensor of the joint motor sends a very small position difference signal. Because. After selecting the output, the output selection unit 127 sends the selected output to the voice recognition unit 129, and the voice recognition unit 129 performs voice recognition.

  In the following, a speech recognition experiment using the speech processing apparatus of this embodiment will be described. In the experiment, the target of speech recognition is a noise signal composed of self noise and background noise mixed with a clear speech utterance. The utterances were conducted using a Japanese database containing 236 words corresponding to four male and female speakers. The position of the speaker is fixed to the front (0 °) throughout the experiment. The recording environment is a room having a size of 4.0 mx 7.0 mx 3.0 m, and the reverberation time is 0.2 seconds. The result of speech recognition is given in terms of word correct rates (WCR).

に基づいて明瞭なスピーチを増幅した。

ここで、Jは、スピーチ活動を伴うセグメントの数であり、s(n)及びd(n)は、n番目の離散したスピーチのサンプル及びノイズのサンプルである。 Before mixing the noise and speech, the segment signal-to-noise ratio shown below to produce an accurate amount of noise and speech energy for various signal-to-noise (signal-to-noise) conditions.

Based on, a clear speech was amplified.

Where J is the number of segments with speech activity and s (n) and d (n) are the nth discrete speech and noise samples.

  FIG. 6 is a diagram illustrating the relationship between the SN ratio (signal-to-noise ratio) and the word accuracy rate (WCR) when there is “head movement” noise. FIG. 6 shows the WCR when the noise reduction unit is not used, the WCR when the multi-channel noise reduction unit is used, and the WCR when the template subtraction noise reduction unit is used. In any S / N ratio, the WCR when the template subtraction noise reduction unit is used is higher than the WCR in the other two cases.

  FIG. 7 is a diagram illustrating the relationship between the SN ratio (signal-to-noise ratio) and the word accuracy rate (WCR) when there is noise of “arm movement”. FIG. 7 shows WCR when the noise reduction unit is not used, WCR when the multi-channel noise reduction unit is used, and WCR when the template subtraction noise reduction unit is used. Except when the S / N ratio is low, the WCR when the template subtraction noise reduction unit is used is lower than the WCR when the multi-channel noise reduction unit is used.

  FIG. 8 is a diagram illustrating the relationship between the SN ratio (signal-to-noise ratio) and the word accuracy rate (WCR) when there is noise of “head movement” and “arm movement”. FIG. 8 shows the WCR when the noise reduction unit is not used, the WCR when the multi-channel noise reduction unit is used, and the WCR when the template subtraction noise reduction unit is used. In any S / N ratio, the WCR when the template subtraction noise reduction unit is used is higher than the WCR in the other two cases.

Table 1 shows the WCR when the noise reduction unit is not used, the WCR when the multi-channel noise reduction unit is used, and the template subtraction reduction unit when there are various noises and the SN ratio is −5 dB. It is the table | surface which showed WCR at the time of using. The unit of WCR in Table 1 is percent.

  As understood from FIGS. 6 to 8 and Table 1, when there is noise of “head movement” and “head and arm movement”, the WCR when the template subtraction noise reduction unit is used is multi-channel. It is higher than WCR when the noise reduction unit is used. On the other hand, when there is noise of “arm movement”, the WCR when using the multi-channel noise reduction unit uses the WCR template subtraction noise reduction unit when using the template subtraction noise reduction unit. Higher than WCR.

  The reason is as follows. The position of the arm is far from the speaker located in front of the robot. Therefore, when there is noise of “arm movement”, noise reduction by sound source separation using a plurality of microphones installed at the head of the robot is performed particularly efficiently in the multi-channel noise reduction unit. Similarly, when there is noise of “leg motion”, noise reduction is efficiently performed by sound source separation using a plurality of microphones.

  On the other hand, when there is noise of “head movement”, the noise due to the head motor propagates in the head with a large reverberation in the vicinity of the microphone. Strong noise sources in the near field of the microphone make the propagation pattern extremely complex. As a result, the noise source deteriorates the separation quality of the sound source separation in the multi-channel noise reduction unit.

  On the other hand, the template subtraction noise reduction unit does not model the noise based on the directional diffusivity, and uses the prediction of the template from the database based on the operation state. Therefore, better results are obtained when there is noise of “head movement” than when using sound source separation.

  From the above results, it is advantageous to use the template subtraction noise reduction unit when there is noise of “head movement” and “head and arm movement”. On the other hand, it is advantageous to use a multi-channel noise reduction unit when there is noise of “arm movement”, “leg movement”, and combinations thereof. Therefore, the output selection unit 127 selects the output of the template subtraction noise reduction unit 100 when “head movement” exists, and outputs the output of the multi-channel noise reduction unit 121 when “head movement” does not exist. Select.

DESCRIPTION OF SYMBOLS 100 ... Template subtraction noise reduction part, 101 ... Operation | movement state acquisition part, 103 ... Sound data acquisition part, 105 ... Database, 107 ... Database search part, 109 ... Template subtraction part, 111 ... Database preparation part

Claims (10)

A sound data processing device for processing sound data including signal sound so as to reduce noise generated by a mechanical device,
An operation state acquisition unit for acquiring an operation state of the mechanical device;
A sound data acquisition unit for acquiring sound data corresponding to the acquired operation state;
A database for storing various operating states of the mechanical device in unit time and corresponding sound data as a template;
A database search unit that searches the database for the operation state closest to the acquired operation state from the database;
A sound data processing apparatus comprising: a template subtraction unit that subtracts sound data of a template in an operation state closest to the acquired operation state from acquired sound data to reduce noise generated by the mechanical device.   A multi-channel noise reduction unit that reduces noise based on multi-channel sound data, and an output that selects either the output of the template subtraction unit or the output of the multi-channel noise reduction unit based on the acquired operating state The sound data processing device according to claim 1, further comprising a selection unit.   The sound data processing device according to claim 1, wherein the mechanical device is a robot.   The sound data processing apparatus according to claim 3, wherein an operation state of the robot is represented by data of an angle, an angular velocity, and an angular acceleration of a joint motor.   The sound data processing device according to claim 4, wherein the template of the operation state closest to the acquired operation state is determined by a distance in a three-dimensional space configured by motor angle, angular velocity, and angular acceleration data.   6. The sound data processing apparatus according to claim 1, wherein the sound data of the template is represented by a frequency spectrum.   The sound data processing device according to claim 1, further comprising a database creation unit that creates the database by collecting various operation states of the mechanical device in unit time and sound data corresponding to the operation states. .   The sound data processing device according to claim 1, wherein the sound data processing device is used for speech recognition. A sound data processing method for processing sound data including a signal sound so as to reduce noise generated by a mechanical device,
Obtaining an operating state of the mechanical device;
Obtaining sound data corresponding to the obtained operating state;
Searching for sound data of a template in an operation state closest to the acquired operation state from a database storing various operation states of the mechanical device in a unit time and corresponding sound data as a template;
Subtracting the sound data of the template in the operation state closest to the acquired operation state from the acquired sound data to obtain an output with reduced noise generated by the mechanical device, and a sound data processing method.   The audio processing according to claim 9, further comprising: selecting one of an output with reduced noise by template subtraction and an output with reduced noise using multi-channel sound data based on the acquired operation state. Method.

Images